This review was based on defined inclusion criteria and underpinned by a search of several databases. Unpublished studies did not appear to be included, which opened the review up to publication bias. Quality was assessed using appropriate tools, but results were not reported in full. A large proportion of the studies did not report blinding and several were subject to selection bias. Two reviewers were involved in selecting studies, data extraction and quality assessment which helped minimise bias and error. Grouping of the studies appeared to have been appropriate.
The authors stated that they used a bivariate model to produce the summary estimates of sensitivity and specificity. However, the pooled estimates of the likelihood ratios did not seem to be derived from these. Therefore, it was unclear whether frequentist meta-analytical techniques, or robust summary receiver operating characteristic curve analyses (which maintain the within-study relationship between the sensitivity and specificity) were used to derive these estimates. Given the heterogeneity of the included studies, the reliability and generalisability of estimates of sensitivity and specificity produced in separate analyses would be uncertain. In addition, results of investigations into between-study heterogeneity were not reported, and it was unclear whether any covariates were included in the bivariate analyses. Although they have been graded by the authors, all of the tests had positive likelihood ratio of over 50, and negative likelihood ratio of 0.3 or lower, so were informative.
Despite some uncertainty regarding the analyses, the overall conclusions are likely to be reliable, although test accuracy may be overestimated to some degree, primarily due to the lack of blinding within the primary studies.